TOGETHER WITH TODAY'S PARTNER
Good morning, {{first_name|there}}. The weights everyone spent two weeks arguing about are now sitting on Hugging Face, free to download — and the switching has already started.
Read time: 3 minutes. Same time, every weekday — rate today's issue at the bottom.
🚀 The Big Story: Kimi K3's weights are live
Moonshot released the full weights for Kimi K3 at 00:00 UTC July 27 — 2.8 trillion parameters, a 1-million-token context window, under a Modified MIT license. It's the largest open-weight model ever published.
The size is the catch: roughly 594GB in native MXFP4, about 1.4TB unquantized. You are not running this on a laptop — most people will reach it through hosted providers.
US companies are already moving. Fortune reports Coinbase, DoorDash and Airbnb are shifting workloads to Chinese open models. Six of the ten most-used models on OpenRouter are now Chinese. DeepSeek-V4-Pro runs $0.87 per million output tokens against roughly $50 for Anthropic's Fable.
The market noticed: Nvidia shed nearly $600B in value after the K3 shock — and China called Washington's sanction threats "AI hegemony" this weekend.
Jason's take: Resist the "everything is free now" read — the sharpest pushback of the weekend argues K3 costs only slightly less per task than OpenAI's top model, because bigger models burn more tokens getting there. Cost per task is the only number that matters, and almost nobody is measuring it. Measure yours this week and you'll make better calls than most CTOs.
⚡ Quick Hits
Nvidia built a security coalition. The Open Secure AI Alliance launched with 40+ members — Microsoft, IBM, CrowdStrike, Cloudflare, Hugging Face, Databricks — after the OpenAI agent breach. Notably absent: OpenAI, Google and Meta.
Hugging Face's CEO went public with demands: release the rogue agents' traces so researchers can study the escape, plus $100M in compute to build defenses.
Nvidia may backstop $250B of OpenAI's 20-year lease on a 10-gigawatt Ohio campus — a ~$500B buildout on former Cold War DOE land.
China's DRAM champion exploded on debut: CXMT rose 472% to a $489B market cap, making it China's most valuable listed company.
Agent-swarm hype met an audit. Bun's much-publicized Rust rewrite is now estimated near $800K with 2,475 open pull requests and no release in 11 weeks — against a headline story of $165K in 11 days.
Nadella warned about an AI bubble on CNN while Microsoft's stock sits down ~24% YoY — and Microsoft is reportedly prioritizing GPUs for its own products over Azure customers.
📡 Trending on X
"Kimi K3 is not cheap" — the contrarian essay climbing Hacker News argues K3 lands only slightly under OpenAI's top model per task, and is expensive next to other Chinese options. It's the most useful thing written about the launch.
The Bun rewrite exposé is the developer story of the week: 1,277 → 2,475 open PRs in 18 days, an estimated 86 days of continuous CI just to merge them. Exhibit A in the agent-productivity debate.
Nadella's bubble comments got clipped and argued over all weekend — a hyperscaler CEO calling for a "democratic AI ecosystem" reads differently when his own capex is the subject.
The productivity paradox essay that hit 143 points: AI's speed gains push people to run more parallel projects instead of finishing fewer — one engineer confessed to 40 concurrent proofs-of-concept before pulling back.
🛠 The Workflow: Test K3 without downloading 1.4TB
You don't need a GPU cluster to find out whether the cheap lane works for you. Thirty minutes:
Skip the download. Reach K3 through a hosted aggregator (OpenRouter lists it alongside your current models) so you can switch with one line, not one weekend.
Pick your three real tasks — the ones you actually repeat weekly. Not benchmarks, not riddles.
Run each on both models and log total tokens spent per finished task, not price per million. This is the number the K3 pricing debate turns on.
Score output 1–5 on your own quality bar, blind if you can manage it.
Move only what wins on both axes — cheaper per task AND equal quality. Everything else stays where it is.
Reply with the word "k3" and I'll send you the cost-per-task comparison sheet.
🧰 Trending Tools
Kimi K3 (open weights) — 2.8T params, 1M context, Modified MIT license on Hugging Face. For teams with real infrastructure or a hosting provider.
OpenRouter — swap models behind one API and compare cost per task. For anyone about to make a switching decision on vibes.
DeepSeek V4-Pro — $0.87 per million output tokens. For high-volume routine work where good-enough is genuinely enough.
CompactifAI — Multiverse's compression shrinks models 80–95% with minimal accuracy loss; they just raised $570M. For edge and on-device deployment.
Grok Build — SpaceXAI open-sourced its coding agent into the new security alliance. For developers who want a free agent to poke at.
📣 Put your brand here. The AI Innovator reaches AI-first operators, creators, and marketers every weekday. Primary sponsorships are now booking.
That's a wrap
Tomorrow: whether anyone outside the hyperscalers can actually serve a 2.8-trillion-parameter model — and what that costs.
Enjoying the format? Share your link — every referral keeps this free:
How was today’s email?
You're reading the 3-minute AI Innovator — same time, every weekday. Hit reply and tell me what you want more of. I read every one.
